Malaysia Weighs Law on Nvidia-Class GPUs at $47B of AI Sites

Malaysia is considering a law that will allow the government to monitor data centre operations, including their use of high-performance chips, Investment, Trade and Industry Minister Johari Abdul Ghani said on Thursday, Reuters reported. Applicants would be expected to disclose the facilities’ operators and locations and the number of graphics processing units (GPUs) being brought in, and the minister said 55 of the 117 approved data centre projects have begun operations, Bernama reported.

The disclosure targets the step that billing data leaves out. As we showed in our analysis of Nvidia’s Singapore filings, Nvidia’s fiscal 2025 10-K put $23,684 million of revenue in Singapore by billing location while shipments to Singapore were under 2% of revenue, so more than $21,074 million billed there went elsewhere, by our arithmetic. In Epoch AI’s data centre dataset, Malaysia’s four sites hold an estimated 1,439,059 H100 equivalents, about two-thirds of the compute it tracks outside the United States, at an estimated $47.38 billion of capital cost, by our arithmetic.

Business Pill · GPUs Explained

An evergreen Business Engineer explainer on why AI runs on GPUs, the chips Malaysia’s proposed law would count. It teaches the general idea only and does not discuss this story.

What the Minister Said

The requirement to disclose operating information, including electricity usage and the number of GPU chips imported, will be embedded in law, the minister said in a speech at the launch of the ministry’s MySaham mobile app, according to Bernama’s report carried by Malay Mail. The government needs to know who the operators and hyperscalers are for each data centre, and monitoring must continue after approval is granted, he said.

At a later press conference, the minister said Malaysia needed to ensure that any chips subject to US restrictions were imported into the country for domestic use rather than for re-export, Reuters reported. “If they want to use this country to trade something that is defined as an illegal transaction, we need to make sure that it will not happen in this country”, he said.

He said the government had no plans to open investigations into Speedmatrix, a Malaysian unit of Singapore-based cloud firm Megaspeed, following allegations that it may have been used to channel Nvidia chips to Chinese users, Reuters reported. “If… they buy the GPUs for their own use in the data centres, then I cannot see what law that we need to use (to investigate them)”, he said.

Megaspeed said in a statement on its website dated Thursday that it took its compliance obligations under US export control laws extremely seriously, Reuters reported. “Megaspeed welcomes appropriate government scrutiny and is committed to demonstrating through concrete actions that it operates responsibly, transparently, and in accordance with applicable US laws”, the company’s chairman and chief executive said.

The Numbers Behind the Disclosure

Of the 117 approved projects, 62, or 53%, had not begun operations, by our arithmetic from the minister’s figures. On power, he said operators must work out their actual capacity needs because suppliers must allocate what is agreed in supply contracts: “For example, if a data centre books a capacity of 100 megawatts (MW) but only uses 50 MW, the remaining 50 MW that are not being used will still be billed”, Bernama quoted him as saying.

Epoch AI’s Frontier Data Centers dataset, in the copy we downloaded on 12 November, lists 96 sites. Its four Malaysian sites, all in Johor, carry an estimated 1,250.7 MW of power and 1,439,059 H100 equivalents, second only to the United States and 4.3 times the dataset’s five Chinese sites, by our arithmetic. In the copy behind our 31 October piece on Malaysia’s tracked compute, the multiple was 4.5.

The largest are DayOne Nusajaya, at an estimated 583,837 H100 equivalents, and DayOne Kempas, at 480,639. Epoch lists Bridge Data Centres MY06 at 221,052, with ByteDance marked as its likely owner and user, and YTL Johor at 153,532. The dataset’s chip types for the four sites are Nvidia’s A100, H100, B200 and B300.

From the Billing Line to the Data Centre Floor

Nvidia’s 10-K for fiscal 2025 said “Customers use Singapore to centralize invoicing while our products are almost always shipped elsewhere”, as we reported. A GPU count filed by each data centre would record what arrives on site, a measure the billing tables do not give.

Malaysia already requires a permit at the border. A ministry statement of 14 July 2025 made all exports, transhipments and transits of high-performance AI chips of US origin subject to a Strategic Trade Permit under Section 12 of the Strategic Trade Act 2010, a catch-all provision that requires notice at least 30 days ahead. By our arithmetic, that was 486 days ago. The statement said the step would close regulatory gaps while Malaysia reviewed adding the chips to its Strategic Items List.

Where Washington Stands

The US Commerce Department is preparing to finalize a rule on remote access to advanced AI chips that would apply expanded controls to seven more countries, including Malaysia, Singapore and Thailand, The Information reported, as set out in our piece on the remote-access rule. Reuters reported the same day that Chinese AI companies are trying to rent Nvidia-powered capacity at data centres outside China, covered in our piece on the capacity rush.

Members of Congress made calls between 9 and 12 November over Nvidia chip sales and Megaspeed, as we counted in our tally of Republican calls. Reuters said Malaysia faces renewed scrutiny over whether it is being used by companies to circumvent US restrictions on advanced AI chips.

The Structural Read

As we read the minister’s remarks, the test he set is domestic use: chips bought for a data centre’s own use, he said, give him no law to investigate under, which is the gap a GPU count and operator disclosure would document.

As we read Epoch’s estimates, the stakes are concentrated: four Johor sites carry about two-thirds of the compute the dataset tracks outside the United States, by our arithmetic, so a disclosure rule would reach a large share of the region’s tracked AI capacity.

Three Implications

FOR DATA CENTRE OPERATORS Operators would disclose their operators, locations, electricity use and the number of GPUs imported, according to the minister.

FOR CHINESE AI DEVELOPERS The minister said Malaysia needed to ensure that chips subject to US restrictions were imported for domestic use rather than re-export, Reuters reported.

FOR WASHINGTON A US rule on remote access, reported by The Information, would apply expanded controls to seven more countries, including Malaysia.

The Business Engineer Lens

This story maps onto the Business Engineer framework AI’s Geopolitical Chokepoint.

The framework argues that the AI supply chain is “creating geopolitical leverage that operates independently of who has the best models”.

As we read it, a host country that counts the GPUs inside its data centres holds part of that leverage, alongside the countries where chips are designed and billed. The framework predates this story and does not discuss it.

What Is Not Established

No draft text of the law has been published, and its timing is not established. Reuters describes a law under consideration, while Bernama’s report says the requirement will be embedded in law. Whether the disclosures would be made public, and whether they would cover facilities already operating, is not stated in the reports we read.

The allegations about Speedmatrix are allegations; the minister said no investigation is planned, and Megaspeed’s statement says it complies with US law. Epoch’s compute and cost figures are estimates. We did not contact the ministry, Megaspeed, Nvidia or the data centre operators named.

The Bottom Line

Malaysia is considering a law that would make data centres disclose their operators, locations, power use and GPU imports, with 55 of 117 approved projects running. By our arithmetic, Epoch’s estimate puts Malaysia’s four tracked sites at about two-thirds of the compute it lists outside the United States, the kind of count Nvidia’s billing-location tables do not provide.